[GLM5] Implement op fusion for GLM5/DSV3.2 (#46876)

This commit is contained in:
Woosuk Kwon
2026-06-28 08:17:39 -07:00
committed by GitHub
parent 5c91039c41
commit 89876b0c54
9 changed files with 1531 additions and 92 deletions
@@ -423,6 +423,7 @@ class MoERunner(MoERunnerInterface):
if (
shared_output is not None
and not self.moe_config.is_sequence_parallel
and not self.moe_config.skip_final_all_reduce
and self._fused_output_is_reduced
):
shared_output = tensor_model_parallel_all_reduce(shared_output)
@@ -445,6 +446,7 @@ class MoERunner(MoERunnerInterface):
# - The MK already reduced the fused output itself.
if (
not self.moe_config.is_sequence_parallel
and not self.moe_config.skip_final_all_reduce
and (self.moe_config.tp_size > 1 or self.moe_config.ep_size > 1)
and not self._fused_output_is_reduced
):